nfs_socket.c

早期freebsd实现

	/*
	 * For reliable protocols, lock against other senders/receivers
	 * in case a reconnect is necessary.
	 * For SOCK_STREAM, first get the Record Mark to find out how much
	 * more there is to get.
	 * We must lock the socket against other receivers
	 * until we have an entire rpc request/reply.
	 */
	if (sotype != SOCK_DGRAM) {
		if (error = nfs_sndlock(&rep->r_nmp->nm_flag, rep))
			return (error);
tryagain:
		/*
		 * Check for fatal errors and resending request.
		 */
		/*
		 * Ugh: If a reconnect attempt just happened, nm_so
		 * would have changed. NULL indicates a failed
		 * attempt that has essentially shut down this
		 * mount point.
		 */
		if (rep->r_mrep || (rep->r_flags & R_SOFTTERM)) {
			nfs_sndunlock(&rep->r_nmp->nm_flag);
			return (EINTR);
		}
		if ((so = rep->r_nmp->nm_so) == NULL) {
			if (error = nfs_reconnect(rep)) {
				nfs_sndunlock(&rep->r_nmp->nm_flag);
				return (error);
			}
			goto tryagain;
		}
		while (rep->r_flags & R_MUSTRESEND) {
			m = m_copym(rep->r_mreq, 0, M_COPYALL, M_WAIT);
			nfsstats.rpcretries++;
			if (error = nfs_send(so, rep->r_nmp->nm_nam, m, rep)) {
				if (error == EINTR || error == ERESTART ||
				    (error = nfs_reconnect(rep))) {
					nfs_sndunlock(&rep->r_nmp->nm_flag);
					return (error);
				}
				goto tryagain;
			}
		}
		nfs_sndunlock(&rep->r_nmp->nm_flag);
		if (sotype == SOCK_STREAM) {
			aio.iov_base = (caddr_t) &len;
			aio.iov_len = sizeof(u_long);
			auio.uio_iov = &aio;
			auio.uio_iovcnt = 1;
			auio.uio_segflg = UIO_SYSSPACE;
			auio.uio_rw = UIO_READ;
			auio.uio_offset = 0;
			auio.uio_resid = sizeof(u_long);
			auio.uio_procp = p;
			do {
			   rcvflg = MSG_WAITALL;
			   error = soreceive(so, (struct mbuf **)0, &auio,
				(struct mbuf **)0, (struct mbuf **)0, &rcvflg);
			   if (error == EWOULDBLOCK && rep) {
				if (rep->r_flags & R_SOFTTERM)
					return (EINTR);
			   }
			} while (error == EWOULDBLOCK);
			if (!error && auio.uio_resid > 0) {
			    log(LOG_INFO,
				 "short receive (%d/%d) from nfs server %s\n",
				 sizeof(u_long) - auio.uio_resid,
				 sizeof(u_long),
				 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
			    error = EPIPE;
			}
			if (error)
				goto errout;
			len = ntohl(len) & ~0x80000000;
			/*
			 * This is SERIOUS! We are out of sync with the sender
			 * and forcing a disconnect/reconnect is all I can do.
			 */
			if (len > NFS_MAXPACKET) {
			    log(LOG_ERR, "%s (%d) from nfs server %s\n",
				"impossible packet length",
				len,
				rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
			    error = EFBIG;
			    goto errout;
			}
			auio.uio_resid = len;
			do {
			    rcvflg = MSG_WAITALL;
			    error =  soreceive(so, (struct mbuf **)0,
				&auio, mp, (struct mbuf **)0, &rcvflg);
			} while (error == EWOULDBLOCK || error == EINTR ||
				 error == ERESTART);
			if (!error && auio.uio_resid > 0) {
			    log(LOG_INFO,
				"short receive (%d/%d) from nfs server %s\n",
				len - auio.uio_resid, len,
				rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
			    error = EPIPE;
			}
		} else {
			/*
			 * NB: Since uio_resid is big, MSG_WAITALL is ignored
			 * and soreceive() will return when it has either a
			 * control msg or a data msg.
			 * We have no use for control msg., but must grab them
			 * and then throw them away so we know what is going
			 * on.
			 */
			auio.uio_resid = len = 100000000; /* Anything Big */
			auio.uio_procp = p;
			do {
			    rcvflg = 0;
			    error =  soreceive(so, (struct mbuf **)0,
				&auio, mp, &control, &rcvflg);
			    if (control)
				m_freem(control);
			    if (error == EWOULDBLOCK && rep) {
				if (rep->r_flags & R_SOFTTERM)
					return (EINTR);
			    }
			} while (error == EWOULDBLOCK ||
				 (!error && *mp == NULL && control));
			if ((rcvflg & MSG_EOR) == 0)
				printf("Egad!!\n");
			if (!error && *mp == NULL)
				error = EPIPE;
			len -= auio.uio_resid;
		}
errout:
		if (error && error != EINTR && error != ERESTART) {
			m_freem(*mp);
			*mp = (struct mbuf *)0;
			if (error != EPIPE)
				log(LOG_INFO,
				    "receive error %d from nfs server %s\n",
				    error,
				 rep->r_nmp->nm_mountp->mnt_stat.f_mntfromname);
			error = nfs_sndlock(&rep->r_nmp->nm_flag, rep);
			if (!error)
				error = nfs_reconnect(rep);
			if (!error)
				goto tryagain;
		}
	} else {
		if ((so = rep->r_nmp->nm_so) == NULL)
			return (EACCES);
		if (so->so_state & SS_ISCONNECTED)
			getnam = (struct mbuf **)0;
		else
			getnam = aname;
		auio.uio_resid = len = 1000000;
		auio.uio_procp = p;
		do {
			rcvflg = 0;
			error =  soreceive(so, getnam, &auio, mp,
				(struct mbuf **)0, &rcvflg);
			if (error == EWOULDBLOCK &&
			    (rep->r_flags & R_SOFTTERM))
				return (EINTR);
		} while (error == EWOULDBLOCK);
		len -= auio.uio_resid;
	}
	if (error) {
		m_freem(*mp);
		*mp = (struct mbuf *)0;
	}
	/*
	 * Search for any mbufs that are not a multiple of 4 bytes long
	 * or with m_data not longword aligned.
	 * These could cause pointer alignment problems, so copy them to
	 * well aligned mbufs.
	 */
	nfs_realign(*mp, 5 * NFSX_UNSIGNED);
	return (error);
}

/*
 * Implement receipt of reply on a socket.
 * We must search through the list of received datagrams matching them
 * with outstanding requests using the xid, until ours is found.
 */
/* ARGSUSED */
nfs_reply(myrep)
	struct nfsreq *myrep;
{
	register struct nfsreq *rep;
	register struct nfsmount *nmp = myrep->r_nmp;
	register long t1;
	struct mbuf *mrep, *nam, *md;
	u_long rxid, *tl;
	caddr_t dpos, cp2;
	int error;

	/*
	 * Loop around until we get our own reply
	 */
	for (;;) {
		/*
		 * Lock against other receivers so that I don't get stuck in
		 * sbwait() after someone else has received my reply for me.
		 * Also necessary for connection based protocols to avoid
		 * race conditions during a reconnect.
		 */
		if (error = nfs_rcvlock(myrep))
			return (error);
		/* Already received, bye bye */
		if (myrep->r_mrep != NULL) {
			nfs_rcvunlock(&nmp->nm_flag);
			return (0);
		}
		/*
		 * Get the next Rpc reply off the socket
		 */
		error = nfs_receive(myrep, &nam, &mrep);
		nfs_rcvunlock(&nmp->nm_flag);
		if (error) {

			/*
			 * Ignore routing errors on connectionless protocols??
			 */
			if (NFSIGNORE_SOERROR(nmp->nm_soflags, error)) {
				nmp->nm_so->so_error = 0;
				if (myrep->r_flags & R_GETONEREP)
					return (0);
				continue;
			}
			return (error);
		}
		if (nam)
			m_freem(nam);
	
		/*
		 * Get the xid and check that it is an rpc reply
		 */
		md = mrep;
		dpos = mtod(md, caddr_t);
		nfsm_dissect(tl, u_long *, 2*NFSX_UNSIGNED);
		rxid = *tl++;
		if (*tl != rpc_reply) {
			if (nmp->nm_flag & NFSMNT_NQNFS) {
				if (nqnfs_callback(nmp, mrep, md, dpos))
					nfsstats.rpcinvalid++;
			} else {
				nfsstats.rpcinvalid++;
				m_freem(mrep);
			}
nfsmout:
			if (myrep->r_flags & R_GETONEREP)
				return (0);
			continue;
		}

		/*
		 * Loop through the request list to match up the reply
		 * Iff no match, just drop the datagram
		 */
		rep = nfsreqh.r_next;
		while (rep != &nfsreqh) {
			if (rep->r_mrep == NULL && rxid == rep->r_xid) {
				/* Found it.. */
				rep->r_mrep = mrep;
				rep->r_md = md;
				rep->r_dpos = dpos;
				if (nfsrtton) {
					struct rttl *rt;

					rt = &nfsrtt.rttl[nfsrtt.pos];
					rt->proc = rep->r_procnum;
					rt->rto = NFS_RTO(nmp, proct[rep->r_procnum]);
					rt->sent = nmp->nm_sent;
					rt->cwnd = nmp->nm_cwnd;
					rt->srtt = nmp->nm_srtt[proct[rep->r_procnum] - 1];
					rt->sdrtt = nmp->nm_sdrtt[proct[rep->r_procnum] - 1];
					rt->fsid = nmp->nm_mountp->mnt_stat.f_fsid;
					rt->tstamp = time;
					if (rep->r_flags & R_TIMING)
						rt->rtt = rep->r_rtt;
					else
						rt->rtt = 1000000;
					nfsrtt.pos = (nfsrtt.pos + 1) % NFSRTTLOGSIZ;
				}
				/*
				 * Update congestion window.
				 * Do the additive increase of
				 * one rpc/rtt.
				 */
				if (nmp->nm_cwnd <= nmp->nm_sent) {
					nmp->nm_cwnd +=
					   (NFS_CWNDSCALE * NFS_CWNDSCALE +
					   (nmp->nm_cwnd >> 1)) / nmp->nm_cwnd;
					if (nmp->nm_cwnd > NFS_MAXCWND)
						nmp->nm_cwnd = NFS_MAXCWND;
				}
				rep->r_flags &= ~R_SENT;
				nmp->nm_sent -= NFS_CWNDSCALE;
				/*
				 * Update rtt using a gain of 0.125 on the mean
				 * and a gain of 0.25 on the deviation.
				 */
				if (rep->r_flags & R_TIMING) {
					/*
					 * Since the timer resolution of
					 * NFS_HZ is so course, it can often
					 * result in r_rtt == 0. Since
					 * r_rtt == N means that the actual
					 * rtt is between N+dt and N+2-dt ticks,
					 * add 1.
					 */
					t1 = rep->r_rtt + 1;
					t1 -= (NFS_SRTT(rep) >> 3);
					NFS_SRTT(rep) += t1;
					if (t1 < 0)
						t1 = -t1;
					t1 -= (NFS_SDRTT(rep) >> 2);
					NFS_SDRTT(rep) += t1;
				}
				nmp->nm_timeouts = 0;
				break;
			}
			rep = rep->r_next;
		}
		/*
		 * If not matched to a request, drop it.
		 * If it's mine, get out.
		 */
		if (rep == &nfsreqh) {
			nfsstats.rpcunexpected++;
			m_freem(mrep);
		} else if (rep == myrep) {
			if (rep->r_mrep == NULL)
				panic("nfsreply nil");
			return (0);
		}
		if (myrep->r_flags & R_GETONEREP)
			return (0);
	}
}

/*
 * nfs_request - goes something like this
 *	- fill in request struct
 *	- links it into list
 *	- calls nfs_send() for first transmit
 *	- calls nfs_receive() to get reply
 *	- break down rpc header and return with nfs reply pointed to
 *	  by mrep or error
 * nb: always frees up mreq mbuf list
 */
nfs_request(vp, mrest, procnum, procp, cred, mrp, mdp, dposp)
	struct vnode *vp;
	struct mbuf *mrest;
	int procnum;
	struct proc *procp;
	struct ucred *cred;
	struct mbuf **mrp;
	struct mbuf **mdp;
	caddr_t *dposp;
{
	register struct mbuf *m, *mrep;
	register struct nfsreq *rep;
	register u_long *tl;
	register int i;
	struct nfsmount *nmp;
	struct mbuf *md, *mheadend;
	struct nfsreq *reph;
	struct nfsnode *np;
	time_t reqtime, waituntil;
	caddr_t dpos, cp2;
	int t1, nqlflag, cachable, s, error = 0, mrest_len, auth_len, auth_type;
	int trylater_delay = NQ_TRYLATERDEL, trylater_cnt = 0, failed_auth = 0;
	u_long xid;
	u_quad_t frev;
	char *auth_str;

	nmp = VFSTONFS(vp->v_mount);
	MALLOC(rep, struct nfsreq *, sizeof(struct nfsreq), M_NFSREQ, M_WAITOK);
	rep->r_nmp = nmp;
	rep->r_vp = vp;
	rep->r_procp = procp;
	rep->r_procnum = procnum;
	i = 0;
	m = mrest;
	while (m) {
		i += m->m_len;
		m = m->m_next;
	}
	mrest_len = i;

	/*
	 * Get the RPC header with authorization.
	 */
kerbauth:
	auth_str = (char *)0;
	if (nmp->nm_flag & NFSMNT_KERB) {
		if (failed_auth) {
			error = nfs_getauth(nmp, rep, cred, &auth_type,
				&auth_str, &auth_len);
			if (error) {
				free((caddr_t)rep, M_NFSREQ);
				m_freem(mrest);
				return (error);
			}
		} else {
			auth_type = RPCAUTH_UNIX;
			auth_len = 5 * NFSX_UNSIGNED;
		}
	} else {
		auth_type = RPCAUTH_UNIX;
		if (cred->cr_ngroups < 1)
			panic("nfsreq nogrps");
		auth_len = ((((cred->cr_ngroups - 1) > nmp->nm_numgrps) ?
			nmp->nm_numgrps : (cred->cr_ngroups - 1)) << 2) +
			5 * NFSX_UNSIGNED;
	}
	m = nfsm_rpchead(cred, (nmp->nm_flag & NFSMNT_NQNFS), procnum,
	     auth_type, auth_len, auth_str, mrest, mrest_len, &mheadend, &xid);
	if (auth_str)
		free(auth_str, M_TEMP);

	/*
	 * For stream protocols, insert a Sun RPC Record Mark.
	 */
	if (nmp->nm_sotype == SOCK_STREAM) {
		M_PREPEND(m, NFSX_UNSIGNED, M_WAIT);
		*mtod(m, u_long *) = htonl(0x80000000 |
			 (m->m_pkthdr.len - NFSX_UNSIGNED));
	}
	rep->r_mreq = m;
	rep->r_xid = xid;
tryagain:
	if (nmp->nm_flag & NFSMNT_SOFT)
		rep->r_retry = nmp->nm_retry;
	else
		rep->r_retry = NFS_MAXREXMIT + 1;	/* past clip limit */
	rep->r_rtt = rep->r_rexmit = 0;
	if (proct[procnum] > 0)
		rep->r_flags = R_TIMING;
	else
		rep->r_flags = 0;
	rep->r_mrep = NULL;

	/*
	 * Do the client side RPC.
	 */
	nfsstats.rpcrequests++;
	/*
	 * Chain request into list of outstanding requests. Be sure
	 * to put it LAST so timer finds oldest requests first.
	 */
	s = splsoftclock();
	reph = &nfsreqh;
	reph->r_prev->r_next = rep;
	rep->r_prev = reph->r_prev;
	reph->r_prev = rep;
	rep->r_next = reph;

	/* Get send time for nqnfs */
	reqtime = time.tv_sec;

	/*
	 * If backing off another request or avoiding congestion, don't
	 * send this one now but let timer do it. If not timing a request,
	 * do it now.
	 */
	if (nmp->nm_so && (nmp->nm_sotype != SOCK_DGRAM ||
		(nmp->nm_flag & NFSMNT_DUMBTIMR) ||
		nmp->nm_sent < nmp->nm_cwnd)) {
		splx(s);
		if (nmp->nm_soflags & PR_CONNREQUIRED)
			error = nfs_sndlock(&nmp->nm_flag, rep);
		if (!error) {
			m = m_copym(m, 0, M_COPYALL, M_WAIT);
			error = nfs_send(nmp->nm_so, nmp->nm_nam, m, rep);
			if (nmp->nm_soflags & PR_CONNREQUIRED)
				nfs_sndunlock(&nmp->nm_flag);
		}
		if (!error && (rep->r_flags & R_MUSTRESEND) == 0) {
			nmp->nm_sent += NFS_CWNDSCALE;
			rep->r_flags |= R_SENT;
		}
	} else {
		splx(s);
		rep->r_rtt = -1;
	}

	/*
	 * Wait for the reply from our send or the timer's.
	 */
	if (!error || error == EPIPE)
		error = nfs_reply(rep);

	/*
	 * RPC done, unlink the request.
	 */
	s = splsoftclock();